48 lines
1.6 KiB
Markdown
48 lines
1.6 KiB
Markdown
---
|
|
license: gpl-3.0
|
|
datasets:
|
|
- Orion-zhen/kto-gutenberg
|
|
language:
|
|
- zh
|
|
- en
|
|
base_model:
|
|
- Orion-zhen/Qwen2.5-7B-Instruct-Uncensored
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Qwen2.5-7B-Gutenberg-KTO
|
|
|
|
This model is fine tuned over gutenberg datasets using kto strategy. It's my first time to use kto strategy, and I'm not sure how the model actually performs.
|
|
|
|
Compared to those large companies which remove accessories such as charger and cables from packages, I have achieved **real** environment protection by **truly** reducing energy consumption, rather than shifting costs to consumers.
|
|
|
|
Checkout GGUF here: [Orion-zhen/Qwen2.5-7B-Gutenberg-KTO-Q6_K-GGUF](https://huggingface.co/Orion-zhen/Qwen2.5-7B-Gutenberg-KTO-Q6_K-GGUF)
|
|
|
|
## Details
|
|
|
|
### Platform
|
|
|
|
~~I randomly grabbed some rubbish from a second-hand market and built a PC~~
|
|
|
|
I carefully selected various dedicated hardwares and constructed an incomparable home server, which I entitled the **Great Server**:
|
|
|
|
- CPU: Intel Core i3-4160
|
|
- Memory: 8G DDR3, single channel
|
|
- GPU: Tesla P4, TDP 75W, boasting its **Eco friendly energy consumption**
|
|
- Disk: 1TB M.2 NVME, PCIe 4.0
|
|
|
|
### Training
|
|
|
|
To practice the **eco-friendly training**, I utilized various methods, including adam-mini, qlora and unsloth, to minimize VRAM and energy usage, as well as accelerating training speed.
|
|
|
|
- dataset: [Orion-zhen/kto-gutenberg](https://huggingface.co/datasets/Orion-zhen/kto-gutenberg)
|
|
- epoch: 2
|
|
- gradient accumulation: 8
|
|
- batch size: 1
|
|
- KTO perf beta: 0.1
|
|
|
|
### Train log
|
|
|
|
|
|
|
|
 |